Computational models for personality prediction based on self-descriptions and evaluative language

Ideja ovoga diplomskoga rada bila je ispitati hipotezu učinkovitijeg izbora komentara s Reddita za predikciju ličnosti. Jedna hipoteza bila je koristiti evaluativni jezik. Autori koriste evaluativni jezik kako bi evaluirali objekte. Budući da literatura ne definira jasno evaluativni jezik i zbog nedostatka istraživanja evaluativnog jezika u području obrade prirodnoga jezika, u radu definiramo evaluativni jezik kao stav autora prema objektu. Na temelju leksikona koji sadrži riječi za izricanje stava filtriramo evaluativne komentare s Reddita. Za filtiranje samoopisa koristimo psihološki upitnik IPIP-NEO. Zatim implementiramo tradicionalne modele strojnog učenja poput stroja potpornih vektora, slučajnih šuma i logističke regresije i dubokih modela poput BERT-a. Pokazujemo na temelju psihološkog modela Big Five da su rezultati za predikciju ličnosti na temelju komentara koji sadrže evaluativni jezik i samoopise daleko bolji od od onih dobivenih slučajnim odabirom komentara.The idea of this work was to test the hypothesis of more efficient selection of comments from social media site Reddit for personality prediction. One hypothesis was the use of evaluative language. Speakers use evaluative language to evaluate objects. Since there is no clear definition of evaluative language in the literature and there is a lack of research on evaluative language in the field of natural language processing, the thesis defines evaluative language as the author's attitude toward the object. Based on the lexicon containing the words expressing the attitude, we filter the Reddit comments containing evaluative language. We use the psychological questionnaire IPIP-NEO to filter the self-description comments. We then implement traditional machine learning models such as the support vector machine, random forest, logistic regression, and deep learning models such as BERT. Using the Big Five psychological model, we show that personality prediction results based on comments containing evaluative language and self-descriptions are far more efficient than those obtained by randomly selecting comments for personality prediction